Multi-matrix Factorization Attention

TL;DR

This paper introduces Multi-matrix Factorization Attention (MFA) and MFA-Key-Reuse (MFA-KR), novel attention architectures that improve model capacity and reduce memory usage under strict Key-Value cache constraints, outperforming existing methods.

Contribution

The paper presents MFA and MFA-KR, new attention mechanisms that enhance capacity and efficiency under cache constraints, a significant advancement over prior multi-head attention variants.

Findings

MFA outperforms MLA under tight KV cache conditions.

MFA-KR reduces memory usage by up to 93.7%.

Both methods maintain competitive performance with standard MHA.

Abstract

We propose novel attention architectures, Multi-matrix Factorization Attention (MFA) and MFA-Key-Reuse (MFA-KR). Existing variants for standard Multi-Head Attention (MHA), including SOTA methods like MLA, fail to maintain as strong performance under stringent Key-Value cache (KV cache) constraints. MFA enhances model capacity by efficiently scaling up both the number and dimension of attention heads through low-rank matrix factorization in the Query-Key (QK) circuit. Extending MFA, MFA-KR further reduces memory requirements by repurposing the key cache as value through value projection re-parameterization. MFA's design enables strong model capacity when working under tight KV cache budget, while MFA-KR is suitable for even harsher KV cache limits with minor performance trade-off. Notably, in our extensive and large-scale experiments, the proposed architecture outperforms MLA and performs comparably to MHA, while reducing KV cache usage by up to 56% and 93.7%, respectively.